Jake Mainwaring

LLMs are everywhere and I believe one of the biggest challenges right now is aligning them to real-world utility Some thoughts from recent work: - Fine-tuning is great, but prompt engineering still wins for many lightweight use cases. - Evaluating with user-facing metrics (not just accuracy or perplexity) changes how you build. - Bias & fairness issues pop up even with apparently clean data. If you’re working with LLMs, what evaluation metrics do you prioritize, and why?

2

1 Comment

Built this graphic to help you understand the impact that Lakebase has inside of the Databricks ecosystem in practical terms. 👉Some positive takeaways: 1) Try it out for your small to medium sized apps today. 2) I'd expect the cost-to-performance to be comparable to some of the non-Databricks options in the market. 3) I'd also expect that the overall total cost of ownership (TCO) of both the application(s) and derivative analytics, ML, and AI work to decrease. 4) Development time and resiliency should move favorably for your applications, specially as Databricks introduces more and more of the Neon features inside Lakebase. 👉 A few callouts: 1) There are a handful of capabilities that are still not supported by Lakebase as some correctly called out on my previous post. 2) The reason why I am not calling these things out (like the lack of compute autoscaling, etc) is because, well, Databricks just acquired Neon, and I expect more of Neon's capabilities to get embedded into Lakebase. 3) I am also expecting that the integration of Neon's capabilities into Lakebase will go a lot faster than the integration of Tabular's capabilities, given the nature of the products. More-so, Databricks had already been partnering with Neon for some time, (hopefully) paving the way for a smoother integration. 👉Bottom-line: Test it out yourself, hopefully it will help decrease your data sprawl, unnecessary ETL, and make development of apps faster!

59

3 Comments

When two $100B gorillas connect, the junior partner has to suck its data throw a very thin straw. Not a good look for either side - Databricks should provide ADBC or ArrowFlight, while Palantir looks like an add-on that Databricks haven't gotten around to implementing yet. JDBC is not how to connect data between big data platforms. Arrow, with ADBC or ArrowFlight, gives columnar to columnar transfers without data serialization/deserialization and without columnar to row-oriented pivots. Hopsworks has had ArrowFlight support since 2023 - https://lnkd.in/dyAAcUSv Databricks and Palantir discuss their new integration in this video - https://lnkd.in/d_c_jSqD

50

8 Comments

Either subscriptions to LLMs will have to be as ubiquitous as insurance where light users subsidize heavy or … a reckoning will need to happen where prices match the actual cost cuz this is unsustainable

2

1 Comment

Document-centric tasks sit at the core of many enterprise, business, and government workflows. Search is important, but the real challenge is going beyond retrieval, enabling systems to reason across documents, verify facts, and handle multi-step information tasks. Great to see a new model from Databricks moving in this direction. I expect we’ll see many more models, including open-weight ones, designed specifically for document-driven workflows, an area with huge potential across enterprise and government use cases.

23

1 Comment

LLMs are being undersold. Most teams treat them like this: retrieve → prompt → respond. That’s not AI strategy. That’s search with extra steps. The breakthroughs happen when you build systems: Evals. Agents. Memory. Tools. Guardrails. Feedback. Prompts don’t scale. Systems do. #AIStrategy #EnterpriseAI #DigitalTransformation #ExecutiveLeadership #FutureOfWork #TechnologyStrategy

4

Venture-backed and hiring ML Engineers? Your only risk is 30 minutes. Here's what that 30 minutes gets you: This week I reviewed about 500 applications for a Senior ML Engineer role. 487 didn't make it. Not because the bar was impossibly high. Because 73% had "ML Engineer" in their title but couldn't explain how gradient descent works. Another 15% had the skills but wanted $450K. The remaining 6% were already knee deep in other processes. (This is why speed is so important!) That's the reality of ML hiring market in 2026. The pipeline looks full. It's actually empty. And your team is burning 40 hours per role finding that out the hard way. We built AlignIQ™ for exactly this. It eliminates 94% of the noise before a human ever touches a resume. What's left gets vetted by real humans who actually understand the ML hiring landscape. You only see candidates who can do the job. One 30-minute call with me. I learn exactly what you need. 72 hours later, screened and vetted ML candidates in your inbox. We're selectively taking on 3 founding clients this week at a rate that won't exist once these spots are filled. As always, you only pay if you hire and we offer flexible terms. If your ML role has been open longer than it should — DM me "ML" and I'll respond today.

2

Data teams: ship a Decisioning Agent v0.1 this week. Use what you already have: - Working ML model - Place your users live, like Slack Add: - a simple agent that connects the two - basic topics: Explain and What-if Outcomes: - Users trust and act - Builders see real demand - ROI on existing models rises Check out my Substack for details and a step-by-step guide on how we do this at Salesforce (Agentforce, Data Cloud, Slack): https://lnkd.in/ePydSDAa

20

This is why I love working at Databricks. Innovation here starts from first principles (and it’s never easy). Our product and R&D teams build “Useful AI” capabilities that enable organizations to generate real alpha, not just add to agentic hype. Add KARL to the list!

40